Abstarct: Violence is one of the serious problems of today's societies that has adverse consequences on the mental and social health of individuals. With the rapid spread of surveillance cameras in social space, identifying and classifying violent behaviors has become one of the fundamental challenges in the field of security and surveillance. This highlights the need to develop efficient and accurate methods for automatically detecting this type of content. In this thesis, an innovative method for detecting violence in surveillance camera images is presented. This method uses a combination of two pre-trained networks, VGG 16 and GoogleNet, to extract image features. The features extracted from these two networks are simultaneously combined and a comprehensive feature vector is created. Then, the random forest algorithm is used to classify this feature vector and detect types of violence. In this work, the UCF Crime databaxse is used to train and evaluate the model. The presented model has achieved an average accuracy of 98.74% in identifying 14 different types of violence including abuse, confinement, arson, assault, robbery, explosion, fighting, regular movies, road accidents, robbery, shooting, shoplifting, theft and property destruction. These results indicate the capability and efficiency of deep learning methods in accurately detecting violence in images and its importance in strengthening surveillance and preventive systems in society.